Guide · Agent skill
How to use the Science Method Investigator agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
The Science skill is an agent skill that applies the scientific method to a problem. Instead of accepting the first explanation, the agent writes a checkable goal, proposes several competing hypotheses that could each be proven wrong, and picks the cheapest checks that separate them. This guide explains what the skill does, how to install it, how to trigger it, and where it does not fit.
Science Method Investigatordanielmiessler/LifeOS · MIT Download the skill Quick start
- 1Download the ZIP. Download the Science skill ZIP from our page.
- 2Unzip into your skills folder. Unzip it into ~/.claude/skills/ for personal use, or into a project's .claude/skills/ folder.
- 3Or upload in the Claude apps. In the Claude apps, open Settings → Capabilities → Skills and upload the skill.
- 4Describe the problem. Ask a question such as why a test is flaky, using natural wording that matches the trigger phrases.
- 5State the goal. Say what fixed looks like in one sentence so the agent can write a checkable GOAL.
- 6Review the hypotheses and first test. Check the competing hypotheses and the test the agent chose first, and why.
What is an agent skill and what does Science make the agent do?
An agent skill is a folder of instructions that an AI agent loads when a request matches the skill's triggers. Science is one such skill, licensed under MIT, and its main file is SKILL.md with supporting files for each working mode.
When Science is active, the agent follows a fixed sequence for questions like why something is happening. It first locks a goal, then names competing explanations, then chooses the test that discriminates among them first. Each verdict has to trace back to evidence checked against criteria set in advance.
- DefineGoal: frames the problem and writes a goal before evidence is collected.
- QuickDiagnosis: a timeboxed debugging pass. The source sets a 15-minute rule for it.
- StructuredInvestigation: the full loop at working grade or research grade, with an auditable record and a stopping decision.
How do you install the Science skill?
You can install the skill in one of three ways. Download the ZIP of the skill from our page and unzip it into your personal skills folder, which is ~/.claude/skills/. Alternatively, unzip it into a project's .claude/skills/ folder so it applies only to that project.
If you use the Claude apps, you can instead upload the skill under Settings → Capabilities → Skills.
- Personal install: unzip into ~/.claude/skills/
- Project install: unzip into the project's .claude/skills/ folder
- Claude apps: Settings → Capabilities → Skills, then upload
How do you trigger the Science skill with natural requests?
You do not need a special command. The skill's description lists trigger phrases, so ordinary requests that match them load it. Phrases in the source include think about, figure out, diagnose, investigate, why is this happening, quick diagnosis, structured investigation, define the goal, what are we trying to achieve, and research-grade investigation.
Requests to repeatedly tune a measured number, such as optimizing a metric, are not handled here. The skill routes those to HillClimb, and multi-angle exploration goes to IterativeDepth.
- "Why is this test flaky?" goes to QuickDiagnosis.
- "Investigate why signups dropped after the release." goes to StructuredInvestigation.
- "Define the goal for the migration before we start." goes to DefineGoal.
- "What might work to cut churn?" goes to HillClimb, because churn is a measured number.
What does an investigation look like?
A quick diagnosis produces a one-sentence GOAL defining what fixed means, at least three hypotheses (H1, H2, H3) each with an observable falsifier, and a note on which test to run first and why. The agent then records whether each hypothesis was CONFIRMED, leading to a fix, or REFUTED, moving to the next one.
A complex investigation produces a Goal Specification with an outcome, a pre-registration timestamp, and an auditable record. The source also says that a goal can be met while a hypothesis is supported, and the two are judged separately.
What does Science need, and does it run code?
The source lists no requirements, does not run scripts, and does not use the network. The skill is made of instruction files that the agent reads.
Science can store your customizations in a folder at <configRoot>/LIFEOS/USER/CUSTOMIZATIONS/SKILLS/Science/, where configRoot is $CLAUDE_CONFIG_DIR when set and ~/.claude otherwise. If that folder exists, its PREFERENCES.md, configurations and resources override the defaults.
- Requirements: none listed in the source
- Runs scripts: no
- Uses network: no
Safety notes and tips for better results
Because the skill runs no scripts and makes no network calls, its own risk is limited to the instructions it gives the agent. Review any commands the agent proposes before you run them on production systems, as you would with any agent output.
To get the most from it, state the problem and what fixed looks like in one sentence. The skill is designed to ask for a checkable goal, so a vague request will produce a slower, less useful start. For obvious five-minute fixes or creative work, the source says the process adds overhead and is not needed.
- Give the observed symptom and what a fixed state looks like.
- Ask for the cheapest test first if you want fast evidence.
- Skip the skill for trivial fixes and creative or generative tasks.
What people use it for
- Tracking down an intermittent test failure by testing competing causes in order of cost.
- Framing a complex business or research question with a goal and criteria before collecting data.
- Writing a pre-registered goal specification for a decision that needs an auditable record.
- Deciding when a repeated tuning request should move to HillClimb instead.
Frequently asked questions
Is the Science agent skill free?
The source states the licence is MIT. It does not list prices, so check the repository for any terms beyond the licence.
What do I need to use the Science skill?
The source lists no requirements. It does not run scripts and does not use the network, so it is made of instruction files the agent reads.
How do I install the Science skill?
Download the ZIP from our page and unzip it into ~/.claude/skills/ for personal use or into a project's .claude/skills/ folder. In the Claude apps, you can upload it under Settings → Capabilities → Skills.
Why is my agent not using the Science skill?
The skill loads when a request matches its trigger phrases, such as why is this happening or investigate. Try wording the request with one of those phrases. Obvious quick fixes and creative tasks are not meant to trigger it.
How is Science different from HillClimb?
Science finds out why something is happening by testing hypotheses. HillClimb improves a measured number through repeated experiments. The source says requests to repeatedly optimise a metric are routed to HillClimb.
When should I use IterativeDepth instead?
Use IterativeDepth for multi-angle exploration of a topic. The Science skill points to it for that purpose and does not handle multi-lens passes.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
Use Science if you debug flaky or intermittent problems or need to frame a complex question before collecting evidence, and you want explanations checked against evidence. Skip it for obvious fixes, creative work, repeated metric tuning (use HillClimb) or broad topic exploration (use IterativeDepth).
Download the skill